This study aims to answer the question of whether Bitcoin's price will rise beyond 5% within the next 7 days by utilizing on-chain, market, and sentiment data from February 2018 to December 2025 by utilizing a novel deep learning model.
Abstract
Bitcoin's future fluctuations are a substantial concern for investments and risk management. Investors and financial institutions require accurate forecasts of these price movements to hedge and optimize portfolios. This study aims to answer the question of whether Bitcoin's price will rise beyond 5% within the next 7 days by utilizing on-chain, market, and sentiment data from February 2018 to December 2025. The proposed model consists of a multi-scale temporal convolutional network with InceptionTCN blocks, CNN channel attention, adaptive average pooling, and a pairwise ranking loss. Dilated convolutions with bottleneck and fusion layers are employed to efficiently capture features over horizons from 1 to 4 days. Given the class imbalance in the dataset, AUC is used instead of accuracy and other classification metrics to reflect the model's performance better. Subsequently, a profit-optimized decision threshold is also applied to align model selection with financial objectives. The proposed model is compared with 5 other baselines: ImprovedTCN_GRU, LSTM, TCN, XGBoost, and Random Forest. Results indicate that the proposed model achieved an AUC of 0.6316 and a profit of 1.703, outperforming all baseline models. Using a novel deep learning model would assist investors in making better financial decisions.
This work attempts to provide a thorough comparative analysis mapping the precise accuracy–efficiency trade-off between Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models under a standardized Grid Search hyperparameter optimization pipeline using a recent Bitcoin closing-price dataset.
A hybrid forecasting framework that integrates sentiment analysis with deep learning to predict Bitcoin’s hourly and daily closing prices and empirical results demonstrate that sentiment-enhanced hybrid models consistently outperform models based solely on technical indicators across RMSE, MAE, MAPE, and R² metrics.
Meltem Kavaklı, Kadriye Filiz Balbal· Balıkesir Üniversitesi Fen B...· 0 citations
A way to estimate the range of Bitcoin prices for the following day by combining Long Short-Term Memory (LSTM) networks with natural language processing (NLP) approaches, showing how sentiment analysis and deep learning architectures can be combined to increase forecasting resilience and interpretability in erratic cryptocurrency markets.
Hajera Fatima, C. B. Jones· International Journal of Eng...· 0 citations
A methodological framework based on LSTM neural networks for comparing the performance of various sets of features in predicting Bitcoin prices and proof that technical indicators have a significant impact on the predictive accuracy of this model, while social media sentiment has a minor impact under the circumstances are brought.
Sedeeq Hasan Banna, Ammar Ahmed Othman, Khaled Al-Raddah· SISTEMASI· 0 citations
Bitcoin is a highly liquid yet volatile crypto asset, making reliable next-day price forecasting challenging. This study develops a Support Vector Regression (SVR) model for predicting the next-day Bitcoin closing price and compares Grid Search tuning with the Whale Optimization Algorithm (WOA) using weighted Time Series Cross-Validation. The dataset comprised 2,620 daily observations from 10 March 2019 to 11 May 2026. Because the final row had no next-day target, it was excluded from scaling, validation, and model selection and retained only for one-step forecasting, leaving 2,619 modelled observations. An 80:20 chronological split produced 2,095 training observations and 524 independent test samples. Grid Search identified C=100, gamma=0.01, and epsilon=0.001 in 8.54 seconds, whereas warm-started WOA identified C=74.21, gamma=0.0075, and epsilon=0.0004 in 220.48 seconds. On the test set, SVR-WOA numerically reduced MAE from USD 1,612.09 to USD 1,551.35, RMSE from USD 2,161.85 to USD 2,112.09, and MAPE from 1.72% to 1.67%, while R² increased from 0.9803 to 0.9812. These differences correspond to improvements of 3.77%, 2.30%, and 3.05%, respectively, but WOA required approximately 25.8 times longer optimization. Because WOA was executed once and evaluated on a single chronological test split, the observed gains should be interpreted as numerical improvements rather than statistically established superiority. The findings therefore demonstrate a trade-off between slightly lower test errors and substantially greater optimization cost.
Wahyudi Wahyudi, Rizky Parlika, Yisti Vita Via· bit-Tech· 0 citations
Bitcoin price forecasting remains difficult because cryptocurrency markets exhibit volatility, nonlinear temporal dependence, and rapid responses to external information. This study examines whether news-derived sentiment improves a Gated Recurrent Unit (GRU) model beyond historical market variables alone. Bitcoin open, high, low, close, and volume data from Binance and Bitcoin-related news from btcnews.com covered January 2020–December 2025. After relevance filtering, 12,332 articles were processed using entity-focused Aspect-Based Sentiment Analysis and FinVADER to generate mean polarity, polarity dispersion, and article-volume features. Two identical two-layer GRU models used the preceding 30-day sequence to predict the next day’s closing price: a baseline with five market variables and an enhanced model with three sentiment features. News was aggregated by publication date without an explicit market-close cut-off; therefore, look-ahead effects cannot be excluded. Data were split chronologically into 70% training, 15% validation, and 15% testing, with scaling fitted only on the training set. In the reported single run, the enhanced model achieved RMSE USD 2,052.65, MAE USD 1,525.67, MAPE 1.51%, and R² 0.9728, compared with USD 2,244.68, USD 1,743.90, 1.70%, and 0.9675 for the baseline. Across five runs, mean RMSE improved from USD 2,402.34 to USD 2,238.46, although variability was greater for the enhanced model. A one-day-ahead Diebold–Mariano test on the selected-run test-error series indicated a significant difference (DM = 3.3605; p = 0.0008), conditional on the reported error construction. Overall, news sentiment provided complementary predictive information, while temporal alignment and initialization sensitivity require further validation.
